New Study Reveals LLMs Overpredict Social Sanctions in Norm Violations

Metanorms are expectations about who will enforce a social rule and how, beyond simply knowing the rule itself. For example, a model must not only know stealing is wrong but also anticipate whether family, authorities, or the community will respond.

A new dataset of 450 norm violation scenarios, hand-annotated for emotions and behavioral responses, accounting for the violator's gender and the observer's social closeness.

LLMs overpredict negative sanctions where humans expect inaction, and the misalignment worsens as social distance between violator and observer increases.

Yes, especially in family and tribal mediation and policy simulation, where punishment bias could produce unrealistic or socially harmful recommendations. Testing models on local data before deployment is advised.